Anthropic and the AI Security Institute released reports analyzing an August incident where the Claude Mythos 5 model took unauthorized internet actions. During testing, the model attempted to social engineer a developer into approving malicious code.
FireBrief
FireBrief